Medicaid enrollment among previously uninsured Americans and associated outcomes by race/ethnicity—United States, 2008‐2014

Objectives To examine the person‐level impact of Medicaid enrollment on costs, utilization, access, and health across previously uninsured racial/ethnic groups. Data Source Medical Expenditure Panel Survey, 2008‐2014. Study Design We pooled multiple 2‐year waves of data to examine the direct impact of Medicaid enrollment among uninsured Americans. We compared changes in outcomes among nonpregnant, uninsured individuals who gained Medicaid (N = 963) to those who remained uninsured (N = 9784) using a difference‐in‐differences analysis. Principal Findings Medicaid enrollment was associated with significant increases in total health care costs and total prescription drug costs and a significant decrease in out‐of‐pocket costs. Among those who gained Medicaid, prescription drug use increased significantly relative to those who remained uninsured. Medicaid enrollment was also associated with a significant increase in reporting a usual source of care, a decrease in foregone care, and significant improvements in severe psychological distress. Changes in total prescription drug costs and total prescription drug fills differed significantly across each racial/ethnic group. Conclusions Among a national sample of uninsured individuals, Medicaid enrollment was associated with substantial favorable changes in out‐of‐pocket costs, prescription drug use, and access to care. Our findings suggest Medicaid is an important tool to reduce insurance‐related disparities among Americans.

WINKELMAN Et AL. evidence stemming from these experiments suggests Medicaid has positive effects on access to care, health, and financial security. [8][9][10][11][12][13][14] For example, Medicaid expansion under the ACA led to an 8.2 percentage point improvement in insurance coverage, 5 a 12.1 percentage point increase in access to primary care, 15 a 3.4 percentage point decrease in self-reported lifetime depression diagnoses among individuals with chronic conditions, 8 and a decrease in unpaid medical bills of $3.4 billion over 2 years. 10 Although the population-level effects of state Medicaid expansions (i.e, average treatment effects) are well documented, less is known about Medicaid's direct impact among people who gain Medicaid after a period of uninsurance (i.e, average treatment effect on the treated). The Oregon Health Insurance Experiment (OHIE), the most rigorous study to date to examine the impact of gaining Medicaid at the individual level, found that uninsured individuals who gained Medicaid in Oregon state had significantly lower levels of depression and out-of-pocket spending and higher levels of prescription medication use than individuals who were not enrolled in Medicaid. [16][17][18] No other contemporary studies have followed individuals who gain Medicaid after a period of uninsurance. Such studies would be helpful to build on the findings of the OHIE and may shed light on whether identified associations are consistent across time, region, and race/ethnicity. These data are critical because they can inform ongoing policy debates regarding the design and funding of Medicaid, as well as efforts to improve racial and ethnic disparities in care. 12,[19][20][21][22] We used a nationally representative panel survey to examine the impact of Medicaid enrollment on disparities in health care costs, access to care, and general health measures among previously uninsured Americans who transitioned onto Medicaid and stratified our analyses by race/ethnicity. Based on findings from the OHIE and populationlevel studies, we hypothesized that Medicaid enrollment would be associated with lower out-of-pocket costs, higher levels of prescription medication use and usual sources of care, and improvements in mental health.

| Data and study population
We used 2008-2014 Medical Expenditure Panel Survey (MEPS) data. Medical Expenditure Panel Survey is a nationally representative survey that compiles demographic, health insurance, health care costs, utilization, and access, and self-reported health data. Medical Expenditure Panel Survey has an overlapping panel design that surveys each respondent five times over a period of 2 years. Therefore, in any given year, half the sample is in their first year and half in their second year. To create our analytic sample, we restricted analyses to respondents who had 2 years of data, were between the ages of 19 and 64 (inclusive) in their first year of MEPS, who were not pregnant in either year, and whose family income was ≤400 percent of the federal poverty level in each year. We excluded pregnant individuals because pregnancy is a categorical eligibility for Medicaid and because patterns of health care in pregnancy are substantively different than for other health circumstances.
Our sample consisted of two groups: (a) those who remained uninsured throughout the 2-year study period and (b) those who gained Medicaid after a period of uninsurance. We defined the latter population as respondents who were uninsured for at least 6 months within their first 9 months in MEPS (Period 1) and had at least 6 months of Medicaid coverage for the remaining 15 months (Period 2). We chose to set our cut-point at 9 months because nearly all individuals who gained Medicaid in our sample would have completed two rounds of surveys while uninsured prior to the fourth quarter of their first year in MEPS, and because this definition is similar to other evaluations of low-income populations who gain Medicaid. 23 To ensure all outcomes derived from round 2 of MEPS occurred in the first 9 months, we excluded individuals who did not complete round 2 by September of their first year in MEPS. Additionally, in sensitivity analyses, we vary each group definition to test the robustness of our results.

| Outcome measures
Health care costs, health care utilization, and self-reported general and mental health were obtained in each of the five MEPS survey rounds, but due to how the sample was created, we did not include values from round 3. We used values from rounds 1 and 2 during Period 1 and rounds 4 and 5 during Period 2 to ensure similar followup time across each period and to allow for a brief washout period between uninsurance and Medicaid enrollment. Access measures and psychological distress were only reported in rounds 2 and 4.

| Health care costs
We examined total health care costs and total out-of-pocket costs for individuals in Period 1 and Period 2, as well as total and out-ofpocket prescription drug costs. Each cost measure was adjusted to 2014 dollars using the Medical Component of the Consumer Price Index. 24 For inpatient, outpatient, and emergency department (ED) visits and prescription drug costs, MEPS collects data from the participating individual and their medical providers. 25

| Health care utilization
We estimated having any ED visit, total number of ED visits per person, any inpatient visit, total number of inpatient visits per person, any prescription drug fill, and total number of prescription drug fills per person. These were obtained through medical provider records.

| Health care access
Several health care access measures were assessed, including a usual source of care, foregone medical care (i.e, "unable to get medical care, tests, or treatments a respondent or a doctor believed to be necessary"), delayed medical care (i.e, "delayed medical care, tests, or treatments a respondent or a doctor believed to be necessary"), or unable to get necessary prescription drugs (i.e, "unable to obtain prescription medicines a respondent or a doctor believed to be necessary"). Each of these outcomes was asked in rounds 2 and 4 and refers to the preceding 12 months.

| Health outcomes
Our final outcome measures included several self-reported health measures: general health (fair or poor health in any survey round for each period), mental health (fair or poor mental health in any survey round for each period), and severe psychological distress (i.e, Kessler index score of 13 or greater). 26

| Covariates
We considered several sociodemographic characteristics that are known to be associated with health insurance status. 6

| Statistical analysis
We first examined whether there were differences between individuals who gained Medicaid and those who remained uninsured by comparing means of baseline sociodemographic characteristics.
Next, for each of our outcomes, we compared baseline values (i.e, in Period 1) to follow-up values (i.e, in Period 2) for individuals who gained Medicaid vs for those who remained uninsured. We also stratified our analysis of Medicaid gainers by race/ethnicity. Due to considerable heterogeneity and low sample size within "Other, non-Hispanic," we excluded this group in stratified analyses. Significance testing of outcome differences between Period 1 and Period 2 was conducted using multivariable linear regression, incorporating the characteristics identified above.
In our final set of analyses, we estimated multivariable linear regression models to assess how gaining Medicaid affected each of our four sets of outcomes relative to remaining uninsured. In each model, we used a difference-in-differences framework, interacting time period with an indicator of Medicaid enrollment, to compare changes in the outcomes for Medicaid gainers to changes among those who remained uninsured between Period 1 and Period 2. Analyses were conducted among the entire sample and also stratified by race/ethnicity. We accounted for the complex survey design in MEPS using svy commands in Stata  In addition to our primary multivariable regression specifications, we ran a series of sensitivity analyses to examine the robustness of our results. First, we compared linear trends for costs, utilization, and self-reported health outcomes in Period 1 among individuals who gained Medicaid and those who remained uninsured. We did not assess Period 1 trends for access measures because data for these outcomes were only collected once during Period 1. Second, to ensure those who gained Medicaid and those who remained uninsured were well matched, we re estimated our difference-in-differences regressions using entropy balancing, an approach that directly reweights the control group to match the means (or other moments) of the treatment group. [28][29][30] We estimated two models using entropy balancing, first weighting with the covariates used in our baseline approach and second, weighting with round 1 and 2 values of the outcomes, using costs when the outcomes were not measured more than once in a period. In both cases, we used the resulting weights to estimate difference-in-differences regressions similar to our baseline analyses. Third, for cost variables, we re-estimated the models using a two-part model. [31][32][33] Finally, we made a variety of modifications to our definitions of both the Medicaid gainer population and the control group, in each case varying the number of months they were either uninsured or had Medicaid coverage.

| Health care costs
Next, we examined how each outcome changed from Period 1 to Period 2 among the entire sample. As indicated in  (Table 3). Increases in total prescription drug costs were statistically significant across racial/ethnic groups.

| Health care utilization
We found no significant changes in ED and inpatient visits among individuals who gained Medicaid compared to those who remained uninsured (

| Health care access
We also observed varying changes in access to care (

| Health outcomes
Individuals who gained Medicaid and those remaining uninsured both reported just under a 2.5 percentage point decrease in the probability of reporting fair or poor health, although this was only statistically significant for those remaining uninsured (

| Difference-in-differences estimates
In addition to examining changes in mean values of each outcome finding that was generally consistent across racial/ethnic groups.
We found no significant differences in changes in ED or inpatient utilization patterns between those gaining Medicaid and those remaining uninsured ( Finally, we found modest changes in health outcomes (Table 4).
We did not find statistically significant differences in changes in selfreported fair/poor general or mental health. However, there was a

| Sensitivity analyses
In our first sensitivity analysis, we examined linear trends in Period 1 using multivariable linear regression models. We found that trends were generally similar for individuals who gained Medicaid and those that remained uninsured. Statistically significant, though quantitatively modest, differences in linear trends during Period 1 were identified for three measures: total costs, any prescription drug fill, and total prescription drug fills (Table S1). Total cost and total prescription fill differences in Period 1 were <25 percent of our difference-in-differences estimate. Therefore, differences in Period 1 are unlikely to explain the large differences observed between Period 1 and Period 2. Differences in Period 1 are also illuminating and suggest individuals who gain Medicaid likely face escalating health costs that may result, through a variety of mechanisms, in enrollment in public health insurance coverage.
Our first entropy-balanced model was weighted on our baseline set of covariates, and our second model was additionally weighted based on round 1 and 2 outcome values, which eliminated Period 1 trend differences (Table S2). Entropy-balanced estimates were substantively similar to each other and to our primary analysis, with one exception. When we weighted on round 1 and 2 outcomes, our difference-in-differences estimates for increases in ED visits and inpatient visits became statistically significant, and the decrease in out-of-pocket spending, while similar in magnitude, was no longer statistically significant.
We also re-estimated cost models using a two-part model. We found smaller, though still statistically significant, increases in total costs and total prescription drug costs, as well as significant reductions in both total out-of-pocket costs and out-of-pocket prescription drug costs (Table S3); these results did not substantively alter our main findings.
In a final sensitivity analysis, we varied the definitions for both study groups. For those gaining Medicaid, we varied the length of uninsured months and time enrolled in Medicaid; for those remaining uninsured, we varied the requirements for the months remaining uninsured. Results were similar to estimates from our primary model specification (Table S4).

| Limitations
Study findings should be considered in light of limitations related to our data source and study design. First, our statistical adjustments may not fully control for selection bias regarding who enrolls in Medicaid. Specifically, somewhat asymmetric trends in Period 1 in comparisons of individuals who remain uninsured vs enroll in Medicaid suggest that uninsured individuals who enroll in Medicaid may do so for reasons related to their need for medical care. While we were unable to fully adjust for these possibilities, we employed entropy balancing in our sensitivity analyses as an additional type of control. Estimates of costs, prescription drug utilization, and access to care from our entropy balancing models were substantively similar to our primary specification.
A second limitation is that only 1 year of data was available after the ACA-sponsored Medicaid expansion occurred in some states (i.e, 2014), and therefore, we could not determine whether gaining Medicaid coverage under the ACA had differential effects compared to gaining Medicaid in prior years. A final limitation is that the access to care outcomes refers to the preceding 12 months. While these measures were generally obtained toward the end of each MEPS data period (i.e, quarter 3 or 4), some responses might refer to previous periods of insurance and bias estimates toward the null. However, we are unaware of other contemporary datasets that could be used to track uninsured populations into Medicaid coverage to examine the outcomes presented in this study at the national level. Healthcare. An earlier version of this manuscript was presented at the 2018 Academy Health Annual Research Meeting in Seattle, WA.